std::chrono 纳秒计时器适用于 MSVC，但不适用于 GCC

Question

I have a simple program that uses chrono for timing that I had ported from MSVC to Code::Blocks. The display of the program shows the delta time from when it was started to 16 decimal places. After getting it to compile, I noticed that the timer was only moving up from the first 6 decimal places. The code remains unchanged, still using std::chrono::high_resolution_clock::now(); for the time, and then to calculate the delta time I use

double localDeltaTime = std::chrono::duration_cast<std::chrono::nanoseconds>(m_EndTime - m_StartTime).count();
localDeltaTime = localDeltaTime / 1000000000.0;

This clearly displays nanosecond timings, yet GCC only seems to do microseconds? Is this a known issue?

Edit:

Here is an MRE,

#include <chrono>
#include <iostream>
#include <iomanip>

int main()
{
    std::chrono::high_resolution_clock::time_point start = std::chrono::high_resolution_clock::now();
    std::chrono::high_resolution_clock::time_point finish = start;

    while (true)
    {
        finish = std::chrono::high_resolution_clock::now();

        long double deltaTime = std::chrono::duration_cast<std::chrono::nanoseconds>(finish - start).count();
        deltaTime /= 1000000000.0;

        std::cout << std::setprecision(25) << deltaTime << std::endl;
    }


    return 0;
}

Answer 1

That's a known MinGW issue #5086 .

It mentions these possible workarounds:

• use std::chrono::steady_clock
• use Boost.chrono
• build your own clock using QueryPerformanceCounter Win32 API

Regarding MSVC:

First of all, on Windows, the best possible user-space timer resolution is 100 ns.

In MSVC system_clock and steady_clock both support this resolution, so you should see 7 decimal digits changing.

But writing to std::cout takes a long time , on the order of 1 ms. So that's the reason why you're seeing large time steps in your version, you're essentially measuring std::cout time.

I rewrote the test to print the minimum change in the time point:

#include <chrono>
#include <iostream>
#include <iomanip>
using namespace std::literals;

template<class Clock>
void runTest() {
    std::cout << typeid(Clock).name() << '\n';
    for (int i = 0; i < 5; i++) {
        auto start = Clock::now();
        for (;;) {
            auto finish = Clock::now();
            if (finish != start) {
                std::cout << std::fixed << std::setprecision(9) << (finish - start) / 1.0s << '\n';
                break;
            }
        }
    }
}

int main() {
    runTest<std::chrono::system_clock>();
    runTest<std::chrono::steady_clock>();
    runTest<std::chrono::high_resolution_clock>();
}

And here's the output I got:

MSVC 19.28 (VS 2019):

struct std::chrono::system_clock
0.000000200
0.000000100
0.000000100
0.000000100
0.000000100
struct std::chrono::steady_clock
0.000000100
0.000000100
0.000000100
0.000000200
0.000000100
struct std::chrono::steady_clock
0.000000200
0.000000100
0.000000100
0.000000100
0.000000100

MinGW-w64 GCC 10.2.0 (Rev1, Built by MSYS2 project):

NSt6chrono3_V212system_clockE
0.000999200
0.000998600
0.000999900
0.001000800
0.000999400
NSt6chrono3_V212steady_clockE
0.000000100
0.000000100
0.000000100
0.000000100
0.000000100
NSt6chrono3_V212system_clockE
0.000999900
0.001001900
0.001006200
0.001016600
0.000980700

So in case of MinGW we can see that at least steady_clock provides 100 ns resolution, but unfortunately high_resolution_clock is an alias of system_clock .